Affiliation:
1. Beijing Advanced Innovation Center for Big Data and Brain Computing, Beijing Key Laboratory for Cooperative Vehicle Infrastructure Systems & Safety Control, School of Transportation Science and Engineering Beihang University Beijing China
2. Beijing Key Laboratory of Multimedia and Intelligent Software Technology, the Faculty of Information Technology Beijing University of Technology Beijing China
3. Civil and Environmental Engineering and H. John Heinz III College Carnegie Mellon University Pittsburgh Pennsylvania USA
4. Discipline of Business Analytics The University of Sydney Business School The University of Sydney Sydney New South Wales Australia
Abstract
AbstractContrastive learning is an increasingly important research direction and has attracted considerable attention in the field of computer vision. It can greatly improve the representativeness of image features through data augmentation, unsupervised learning, and pre‐trained models. However, in the field of traffic flow forecasting, most graph‐based models focus on the construct of spatial–temporal relationships between road segments and ignore the use of temporal data augmentation and pre‐trained models, which can improve the representation ability of the forecasting model. Therefore, in this work, contrastive learning are used to expand the distribution of sequence samples and improve the quality and generalization of forecasting models. Based on this, a novel forecasting model called contrastive learning based on multi graph convolution network (CLMGCN) is proposed, which is combined with four components: multi graph convolution network, which learns the spatial–temporal feature of the input traffic data; temporal data augmentation, which obtains the augmentation data of the input traffic data; contrastive learning, which achieves the pre‐training phase and improve the quality of output feature of multi graph convolution network; output block, which utilizes the enhanced output feature of multi graph convolution network for predicting the future traffic data. Finally, by the experimental results of four public traffic flow datasets, it can be shown that CLMGCN achieves higher traffic forecasting accuracy with lower model complexity.
Funder
National Key Research and Development Program of China
National Natural Science Foundation of China
China Postdoctoral Science Foundation
Natural Science Foundation of Beijing Municipality
Publisher
Institution of Engineering and Technology (IET)
Subject
Law,Mechanical Engineering,General Environmental Science,Transportation